CN113111744B - Vein identification method based on time domain short-time and long-time feature fusion - Google Patents

Vein identification method based on time domain short-time and long-time feature fusion Download PDF

Info

Publication number
CN113111744B
CN113111744B CN202110334529.7A CN202110334529A CN113111744B CN 113111744 B CN113111744 B CN 113111744B CN 202110334529 A CN202110334529 A CN 202110334529A CN 113111744 B CN113111744 B CN 113111744B
Authority
CN
China
Prior art keywords
time
long
short
layer
neural network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202110334529.7A
Other languages
Chinese (zh)
Other versions
CN113111744A (en
Inventor
赵明剑
谢斯雅
韦岗
曹燕
王一歌
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
South China University of Technology SCUT
Original Assignee
South China University of Technology SCUT
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by South China University of Technology SCUT filed Critical South China University of Technology SCUT
Priority to CN202110334529.7A priority Critical patent/CN113111744B/en
Publication of CN113111744A publication Critical patent/CN113111744A/en
Application granted granted Critical
Publication of CN113111744B publication Critical patent/CN113111744B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G06F2218/04—Denoising
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction

Abstract

The invention discloses a vein identification method based on time domain short-time and long-time feature fusion, which obtains short-time features and long-time features based on a time domain period by designing a multi-dimensional vein information extraction strategy. Wherein the short-term features are composed of a primary wave envelope, a secondary wave envelope, and a primary and secondary wave envelope, and the long-term features are multi-cycle features corresponding to a respiratory cycle. The invention combines the long and short characteristics based on the time domain period with the single-period pulse wave signals with regular starting points, and inputs the signals into a two-way branch neural network for characteristic fusion recognition. The vein recognition method provided by the invention is used for identity recognition, has the characteristics of high accuracy, strong reliability and strong generalization capability, and is suitable for various practical scenes of identity verification.

Description

Vein identification method based on time domain short-time and long-time feature fusion
Technical Field
The invention relates to the technical field of artificial intelligence machine learning, in particular to a vein identification method based on time domain short-time and long-time feature fusion.
Background
With the explosive growth of information services and digitization systems, password management becomes increasingly difficult in real life. In recent years, scientists have used biometric identification techniques to model personalized features of people, replacing passwords with the uniqueness of people. However, common biometric technologies, including fingerprints, irises, geometric shapes of hands, faces and other biometric information, are easily attacked, and most biometric technologies require expensive acquisition and living body detection, and the extraction cost is high. Meanwhile, new photoplethysmography (PPG) and electrocardiography (ecg) are non-invasive techniques currently used for cardiovascular diagnosis, and have been widely studied and utilized as data sources for biometric identification technology in recent years. PPG is a simple, low-cost optical technique that detects blood volume changes in blood vessels by measuring the skin surface. PPG sensors have been applied today in many different wearable devices. Unlike the electrocardiogram, the signal acquisition of the PPG sensor only needs to be done on one side of the body, which is more easy to use and practical.
At present, the extraction of the vein information is mostly based on a single-dimensional time domain characteristic or a frequency domain characteristic. The single-dimensional time domain feature extraction is divided into two types, wherein one type is to take the whole pulse wave signal as input, and the other type is to input the signal after being segmented according to the period. The former easily causes high complexity of the model, large parameter quantity and needs to process the condition that the starting points of different pulse wave signals are not aligned, and the latter easily causes the model to pay excessive attention to the feature extraction in the period after segmentation, and ignores the relationship in different periods.
Disclosure of Invention
The invention aims to solve the defects in the prior art and provides a vein identification method based on time domain short-time and long-time feature fusion.
The purpose of the invention can be achieved by adopting the following technical scheme:
a vein recognition method based on time domain short-time and long-time feature fusion comprises the following steps:
s1, a pulse wave database is built by self, pulse wave signals stored in the pulse wave database are preprocessed through wavelet de-noising, smooth filtering and data normalization, and then short-time features, long-time features and single-period pulse wave signals with regular starting points are obtained through a multi-dimensional vein information extraction strategy, wherein the short-time features comprise a main wave envelope, a secondary wave envelope and a main and secondary wave envelope, and the long-time features are multi-period long-time features corresponding to a respiratory period;
s2, constructing a two-way branch neural network, wherein the two-way branch neural network comprises a short-time signal processing branch and a long-time signal processing branch, constructing a short-time signal processing branch of a convolution layer plus a nonlinear activation function, and inputting short-time characteristics into the short-time signal processing branch; constructing a long-time signal processing branch of a convolution-circulation structure, and inputting the long-time characteristics and the single-period pulse wave signals with regular starting points into the long-time signal processing branch;
and S3, performing feature fusion on the output result of the two-way branch neural network in the step S2 to obtain a vein recognition result.
Further, the establishment process of the pulse wave database is as follows:
through the pulse oximeter, under the condition of a laboratory, data are collected at wrists and fingers of different users, so that pulse wave signals of the different users are obtained.
Further, the process of preprocessing the pulse wave signal in step S1 is as follows:
filtering high-frequency noise in the pulse wave signals through wavelet denoising and smooth filtering, and reserving a low-frequency part; and then removing the baseline drift, and performing data normalization to obtain the processed pulse wave signal.
Further, the feature extraction process of the short-term feature and the long-term feature in step S1 is as follows:
connecting the main wave peak values of all pulse wave signals, and performing data enhancement by using a cubic spline interpolation mode to obtain a main wave envelope;
connecting the secondary wave peak values of all pulse wave signals, and performing data enhancement by using a cubic spline interpolation mode to obtain secondary wave envelopes;
connecting the primary-secondary wave peak values of each group of all pulse wave signals, and performing data enhancement by using a cubic spline interpolation mode to obtain primary and secondary wave envelopes;
the short-time feature extraction can enrich feature content, make up for the deficiency of trend information after segmenting the pulse wave signals, and experiments show that the auxiliary features can improve the identification accuracy rate of the single-period pulse wave signals.
The signal of the period of increasing the number of points in the pulse wave signal period is the signal during inspiration, the signal of the period of decreasing the number of points in the period is the signal during expiration, and the sequence of increasing and then decreasing the number of points in each continuous period of the pulse wave signal is used as the multi-period long-term characteristic.
Obtaining a sequence of the main waves by calculating a maximum value, then regarding a time interval between every two main waves as a period, and performing data enhancement by using a linear interpolation method, wherein after interpolation, each period has the same starting point and the same number of points, and the signals are the single-period pulse wave signals with regular starting points.
Furthermore, the network structure of the short-time signal processing branch comprises an input layer, a one-dimensional convolutional neural network layer and a nonlinear layer which are sequentially connected; wherein the nonlinear layer uses a ReLu function.
Furthermore, the long-term signal processing branch adopts a convolution-cyclic neural network, and the network structure of the long-term signal processing branch comprises an input layer, a two-dimensional convolution neural network layer, a cyclic neural network layer and a nonlinear layer which are sequentially connected, wherein the nonlinear layer uses a ReLu function. Compared with the feature extraction of the pulse wave signals by using a common one-dimensional convolution neural network, the two-dimensional convolution neural network can be used for extracting the time-invariant features in the pulse wave signals, and experiments show that the two-dimensional convolution neural network is used for replacing the one-dimensional convolution neural network, so that the effect of the model is greatly improved.
Further, the process of feature fusion in step S3 is as follows:
the output of the short-time signal processing branch and the output of the long-time signal processing branch in the double-path branch neural network are connected in series and sent into a full-connection neural network, and the network structure of the full-connection neural network comprises an input layer, a full-connection layer, a nonlinear layer 1, a full-connection layer, a nonlinear layer 2 and a full-connection layer which are sequentially connected in sequenceA tie layer, a nonlinear layer 3; wherein, the nonlinear layer 1 and the nonlinear layer 2 use ReLu function, and the nonlinear layer 3 uses Sigmoid function; taking the output of the last layer of the fully-connected neural network as a predicted value p for obtaining a sample i i When p is i When the pulse width is more than or equal to 0.5, the two pieces of pulse line information of the sample i are considered to belong to the same person, and when p is greater than or equal to 0.5 i When the value is less than 0.5, the two pieces of vein information of the sample i are not considered to belong to the same person.
Further, the two-way branch neural network is trained by using a cross entropy loss function L, which is as follows:
Figure BDA0002997583920000041
wherein y is i Representing the true value, p, of the ith sample i And N is the number of samples.
Further, the implementation of the one-dimensional convolutional neural network layer in the short-time signal processing branch is as follows:
Figure BDA0002997583920000042
wherein, the first and the second end of the pipe are connected with each other,
Figure BDA0002997583920000043
representing the output of the nth channel after convolution of the kth short-time feature, S _ C in The input dimension representing the short-term feature,
Figure BDA0002997583920000044
a parameter, S _ in, corresponding to the kth short-time feature in the output nth channel k Representing the input of the kth short-time feature.
Further, the two-dimensional convolutional neural network layer in the long-term signal processing branch is implemented as follows:
Figure BDA0002997583920000045
wherein the content of the first and second substances,
Figure BDA0002997583920000046
represents the output of the nth channel after the k long-term feature is convoluted,
Figure BDA0002997583920000047
the input dimension representing the kth long-term feature,
Figure BDA0002997583920000048
a parameter L _ in representing the correspondence of the kth long-term feature in the output nth channel k An input representing a k-th long-term feature;
the implementation of the recurrent neural network layer in the long-term signal processing branch is as follows:
i t ＝σ(W i x t +U i h t-1 ) (4)
f t ＝σ(W f x t +U f h t-1 ) (5)
o t ＝σ(W o x t +U o h t-1 ) (6)
Figure BDA0002997583920000051
Figure BDA0002997583920000052
Figure BDA0002997583920000053
wherein the function
Figure BDA0002997583920000054
Represent
Figure BDA0002997583920000055
i t 、f t 、o t An input gate, a forgetting gate and an output gate at the t-th moment, x t Indicates the input at the t-th time, h t-1 Representing the hidden gate output at time t-1,
Figure BDA0002997583920000056
candidate memory cell representing the t-th time, c t The memory cell at the t-th time is shown,
Figure BDA0002997583920000057
representing dot multiplication, where W and U are trainable matrices, W i ，W f ，W o ，W c Respectively show the input x in the input gate, the forgetting gate, the output gate and the memory unit t Corresponding parameter matrix, U i 、U f 、U o 、U c Respectively shown in the input gate, the forgetting gate, the output gate and the memory unit, and the output h of the hidden gate t-1 The corresponding parameter matrix.
Compared with the prior art, the invention has the following advantages and effects:
(1) Compared with the manual feature extraction, the method can automatically extract the features from the sequence.
(2) The invention uses multidimensional characteristics, simultaneously considers the characteristics in the period and the trend characteristics outside the period, and the complementary characteristic composition can provide more comprehensive information clues, has more excellent effect and stronger robustness.
(3) The invention also focuses on the problem of alignment between sequences and within sequences, which makes the invention achieve better effect under the balance of effect and parameter quantity.
Drawings
Fig. 1 is a flowchart of a vein identification method based on time domain short-time and long-time feature fusion disclosed in the embodiment of the present invention;
FIG. 2 is a flow diagram of a short-time signal processing branch in an embodiment of the present invention;
fig. 3 is a flow chart of a long-term signal processing branch according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Examples
The embodiment discloses a vein identification method based on time domain short-time and long-time feature fusion, as shown in fig. 1, the specific steps are as follows:
s1, a pulse wave database is built, pulse wave signals stored in the pulse wave database are preprocessed through wavelet denoising, smoothing filtering and data normalization, and then short-time features, long-time features and single-period pulse wave signals with regular starting points are obtained through a multi-dimensional pulse texture information extraction strategy, wherein the short-time features comprise a main wave envelope, a secondary wave envelope and a main and secondary wave envelope, and the long-time features are multi-period long-time features corresponding to a respiratory period;
the establishment process of the pulse wave database is as follows:
through the pulse oximeter, in a laboratory environment, data acquisition is performed at wrists and fingers of different users, so that pulse wave signals of the different users are obtained. The frequency of sampling of each set of pulse wave signals is 500hz, and 10221 points are acquired in total each time. After continuously testing for one month, signals of 37 persons are collected, and 85002 effective data are finally obtained by filtering, wherein each data comprises 10221 points and is about 22 periodic pulse wave signals.
Wherein, the pulse wave signal preprocessing process in the step S1 is as follows:
filtering high-frequency noise in the pulse wave signals by wavelet denoising and smoothing filtering, and reserving a low-frequency part; and then removing the baseline drift, and carrying out data normalization to obtain the processed pulse wave signal.
The feature extraction process of the short-term features and the long-term features in the step S1 is as follows:
connecting the main wave peak values of all pulse wave signals, performing data enhancement by using a cubic spline interpolation mode, and extending the main wave envelope to 100 points to obtain the main wave envelope;
connecting the secondary wave peak values of all pulse wave signals, performing data enhancement by using a cubic spline interpolation mode, and extending the secondary wave envelope to 100 points to obtain the secondary wave envelope;
connecting the primary-secondary wave peak values of each group of all pulse wave signals, performing data enhancement by using a cubic spline interpolation mode, and extending primary and secondary wave envelopes to 100 points to obtain primary and secondary wave envelopes;
the signal of the point number increasing period in the pulse wave signal period is the signal during inspiration, the signal of the point number decreasing period in the period is the signal during expiration, and the sequence of increasing and then decreasing the point number in each continuous period of the pulse wave signal is used as the multi-period long-time characteristic. The sequence was extended to 2500 points using linear interpolation as a means of data enhancement. In order to utilize matrix parallelization acceleration at a hardware end, zero filling is carried out on sequences with maximum length of short multi-cycle long-term features;
obtaining a sequence of the main waves by calculating a maximum value, then regarding a time interval between every two main waves as a period, and performing data enhancement by using a linear interpolation method to prolong the sequence to 500 points. After interpolation, the signal with the same starting point and the same number of points in each period is the single-period pulse wave signal with the regulated starting point. In order to accelerate parallelization of the matrix at a hardware end, zero filling is carried out on a sequence with the maximum length of the single-period pulse wave signals with the deficiency of the normalized starting points. And cutting the start-point-structured monocycle wave signals into a matrix of N x 500, wherein N represents the number of cycles contained in the extended start-point-structured monocycle wave signals.
S2, constructing a two-way branch neural network, wherein the two-way branch neural network comprises a short-time signal processing branch and a long-time signal processing branch, constructing a short-time signal processing branch of a convolution layer plus a nonlinear activation function, and inputting short-time characteristics into the short-time signal processing branch; constructing a long-time signal processing branch of a convolution-circulation structure, and inputting the long-time characteristics and the single-period pulse wave signals with regular starting points into the long-time signal processing branch;
and S3, performing feature fusion on the output result of the two-way branch neural network in the step S2 to obtain a vein recognition result.
Wherein, the process of feature fusion in the step S3 is as follows:
the output obtained by the two-way branch neural network is connected in series and sent into a full-connection neural network, and the network structure of the full-connection neural network comprises an input layer, a full-connection layer, a nonlinear layer 1, a full-connection layer, a nonlinear layer 2, a full-connection layer and a nonlinear layer 3 which are sequentially connected; wherein, the nonlinear layer 1 and the nonlinear layer 2 use ReLu function, and the nonlinear layer 3 uses Sigmoid function; taking the output of the last layer of the fully-connected neural network as a predicted value p of a sample i i When p is i When the pulse width is more than or equal to 0.5, the two pieces of pulse line information of the sample i are considered to belong to the same person, and when p is greater than or equal to 0.5 i When the value is less than 0.5, the two pieces of vein information of the sample i are not considered to belong to the same person.
The two-way branch neural network is trained by using a cross entropy loss function L, which specifically comprises the following steps:
Figure BDA0002997583920000081
wherein y is i Representing the true value, p, of the ith sample i And N is the number of samples.
The two-way branch neural network in step S2 refers to a short-time signal processing branch and a long-time signal processing branch.
The specific flow of the training of the short-time signal processing branch is shown in fig. 2, the specific number of layers of the short-time signal processing branch is constructed as shown in table 1, and the construction method of the short-time signal processing branch is to input short-time characteristics including a main wave envelope, a secondary wave envelope and a main and secondary wave envelope into the short-time signal processing branch; the network structure of the short-time signal processing branch comprises an input layer, a one-dimensional convolutional neural network layer and a nonlinear layer which are sequentially connected; wherein the nonlinear layer uses a ReLu function.
The implementation of the one-dimensional convolutional neural network layer in the short-time signal processing branch is as follows:
Figure BDA0002997583920000082
wherein the content of the first and second substances,
Figure BDA0002997583920000083
representing the output of the nth channel after convolution of the kth short-time feature, S _ C in An input dimension representing a short-term feature,
Figure BDA0002997583920000084
a parameter, S _ in, corresponding to the kth short-time feature in the nth channel of the output k Representing the input of the kth short-time feature.
TABLE 1 parameter tables of respective layers
Figure BDA0002997583920000085
Figure BDA0002997583920000091
The specific flow of the long-term signal processing branch training is shown in fig. 3, the long-term signal processing branch is constructed as shown in table 2, and the long-term signal processing branch construction method is to input the long-term signal processing branch into the multi-cycle long-term feature and the single-cycle pulse wave signal with regular starting points; the long-time signal processing branch adopts a convolution-circulation neural network, and the network structure of the long-time signal processing branch comprises an input layer, a two-dimensional convolution neural network layer, a circulation neural network layer and a nonlinear layer which are sequentially connected, wherein the nonlinear layer uses a ReLu function.
The two-dimensional convolutional neural network layer in the long-term signal processing branch is realized as follows:
Figure BDA0002997583920000092
wherein
Figure BDA0002997583920000093
Represents the output of the nth channel after the k long-term feature is convoluted,
Figure BDA0002997583920000094
the input dimension representing the kth long-term feature,
Figure BDA0002997583920000095
a parameter, L _ in, corresponding to the kth long-term feature in the output nth channel k Representing the input of the kth long-term feature.
The implementation of the recurrent neural network layer in the long-term signal processing branch is as follows:
i t ＝σ(W i x t +U i h t-1 ) (4)
f t ＝σ(W f x t +U f h t-1 ) (5)
o t ＝σ(W o x t +U o h t-1 ) (6)
Figure BDA0002997583920000096
Figure BDA0002997583920000101
Figure BDA0002997583920000102
wherein
Figure BDA0002997583920000103
To represent
Figure BDA0002997583920000104
i t 、f t 、o t An input gate, a forgetting gate and an output gate at the t-th moment, x t Indicates the input at the t-th time, h t-1 Representing the hidden gate output at time t-1,
Figure BDA0002997583920000105
candidate memory cell representing the t-th time, c t The memory cell at the t-th time is shown,
Figure BDA0002997583920000107
representing dot multiplication, where W and U are trainable matrices, W i ，W f ，W o ，W c Respectively show the input x in the input gate, the forgetting gate, the output gate and the memory unit t Corresponding parameter matrix, U i ，U f ，U o ，U c Respectively shown in the input gate, the forgetting gate, the output gate and the memory unit, and the output h of the hidden gate t-1 The corresponding parameter matrix.
Table 2.N shows a table of the period numbers of the input signals
Figure BDA0002997583920000106
The specific number of layers of the fully-connected neural network in the step S3 is shown in table 3.
Table 3.M shows the long time signal branch and the short time signal branch output concatenation dimension table
Network layer Detailed description of the preferred embodiment Characteristic dimension
Input layer Is free of 1*M
Full connection layer The weight matrix is M256 1*256
Non-linear layer 1 ReLu function 1*256
Full connection layer The weight matrix is 256 × 128 1*128
Non-linear layer 2 Relu function 1*128
Full connection layer The weight matrix is 128 x 1 1*1
Non-linear layer 3 Sigmoid function 1
Output layer Is free of 1
The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such modifications are intended to be included in the scope of the present invention.

Claims (6)

1. A vein identification method based on time domain short-time and long-time feature fusion is characterized by comprising the following steps:
s1, a pulse wave database is built, pulse wave signals stored in the pulse wave database are preprocessed through wavelet denoising, smoothing filtering and data normalization, and then short-time features, long-time features and single-period pulse wave signals with regular starting points are obtained through a multi-dimensional pulse texture information extraction strategy, wherein the short-time features comprise a main wave envelope, a secondary wave envelope and a main and secondary wave envelope, and the long-time features are multi-period long-time features corresponding to a respiratory period;
the feature extraction process of the short-term feature and the long-term feature in the step S1 is as follows:
connecting the main wave peak values of all pulse wave signals, and performing data enhancement by using a cubic spline interpolation mode to obtain a main wave envelope;
connecting the secondary wave peak values of all pulse wave signals, and performing data enhancement by using a cubic spline interpolation mode to obtain secondary wave envelopes;
connecting the primary-secondary wave peak values of each group of all pulse wave signals, and performing data enhancement by using a cubic spline interpolation mode to obtain primary and secondary wave envelopes;
the signal of the point number increasing period in the pulse wave signal period is the signal during inspiration, the signal of the point number decreasing period in the period is the signal during expiration, and the sequence of increasing and then decreasing the point number in each continuous period of the pulse wave signal is used as the multi-period long-time characteristic;
obtaining a sequence of main waves by calculating a maximum value, then regarding a time interval between every two main waves as a period, and performing data enhancement by using a linear interpolation method, wherein a signal with the same starting point and the same point number in each period after interpolation is a single-period pulse wave signal with a regular starting point;
s2, constructing a two-way branch neural network, wherein the two-way branch neural network comprises a short-time signal processing branch and a long-time signal processing branch, constructing a short-time signal processing branch of a convolution layer plus a nonlinear activation function, and inputting short-time characteristics into the short-time signal processing branch; constructing a long-time signal processing branch of a convolution-circulation structure, and inputting the long-time characteristics and the single-period pulse wave signals with regular starting points into the long-time signal processing branch; the network structure of the short-time signal processing branch comprises an input layer, a one-dimensional convolutional neural network layer and a nonlinear layer which are sequentially connected; wherein the nonlinear layer uses a ReLu function; the long-term signal processing branch adopts a convolution-cyclic neural network, and the network structure of the long-term signal processing branch comprises an input layer, a two-dimensional convolution neural network layer, a cyclic neural network layer and a nonlinear layer which are sequentially connected, wherein the nonlinear layer uses a ReLu function;
s3, performing feature fusion on the output result of the two-way branch neural network in the step S2 to obtain a vein recognition result, wherein the process of the feature fusion is as follows:
the method comprises the following steps of (1) serially connecting the outputs of a short-term signal processing branch and a long-term signal processing branch in a two-way branched neural network, and sending the outputs into a fully-connected neural network, wherein the network structure of the fully-connected neural network comprises an input layer, a fully-connected layer, a nonlinear layer 1, a fully-connected layer, a nonlinear layer 2, a fully-connected layer and a nonlinear layer 3 which are sequentially connected; wherein, the nonlinear layer 1 and the nonlinear layer 2 use ReLu function, and the nonlinear layer 3 uses Sigmoid function; taking the output of the last layer of the fully-connected neural network as a predicted value p for obtaining a sample i i When p is i When the pulse width is more than or equal to 0.5, the two pieces of pulse line information of the sample i are considered to belong to the same person, and when p is greater than or equal to 0.5 i When the value is less than 0.5, the two pieces of vein information of the sample i are not considered to belong to the same person.
2. The method for pulse texture recognition based on time domain short-term and long-term feature fusion as claimed in claim 1, wherein the establishment of the pulse wave database is as follows:
through the pulse oximeter, data acquisition is performed at wrists and fingers of different users under laboratory conditions, so that pulse wave signals of different users are obtained.
3. The method for pulse-vein recognition based on time-domain short-time and long-time feature fusion as claimed in claim 1, wherein the pulse wave signal preprocessing in step S1 comprises:
filtering high-frequency noise in the pulse wave signals by wavelet denoising and smoothing filtering, and reserving a low-frequency part; and then removing the baseline drift, and carrying out data normalization to obtain the processed pulse wave signal.
4. The method for recognizing the vein based on the time-domain short-time and long-time feature fusion as claimed in claim 1, wherein the two-way branch neural network is trained by using a cross entropy loss function L as follows:
Figure FDA0003990433170000031
wherein y is i Representing the true value, p, of the ith sample i And N is the number of samples.
5. The method for recognizing the vein based on the fusion of the short-time characteristic and the long-time characteristic of the time domain according to claim 1, wherein the implementation of the one-dimensional convolutional neural network layer in the short-time signal processing branch is as follows:
Figure FDA0003990433170000032
wherein the content of the first and second substances,
Figure FDA0003990433170000033
representing the kth short-time feature by convolutionOutput of the last n channel, S _ C in The input dimension representing the short-term feature,
Figure FDA0003990433170000034
a parameter, S _ in, corresponding to the kth short-time feature in the output nth channel k Representing the input of the kth short-time feature.
6. The method for recognizing the veins based on the time-domain short-time and long-time feature fusion of claim 1, wherein the two-dimensional convolutional neural network layer in the long-time signal processing branch is implemented as follows:
Figure FDA0003990433170000035
wherein, the first and the second end of the pipe are connected with each other,
Figure FDA0003990433170000041
represents the output of the nth channel after the k long-term feature is convoluted,
Figure FDA0003990433170000042
the input dimension representing the kth long-term feature,
Figure FDA0003990433170000043
a parameter, L _ in, corresponding to the kth long-term feature in the output nth channel k An input representing a kth long-term feature;
the implementation of the recurrent neural network layer in the long-term signal processing branch is as follows:
i t ＝σ(W i x t +U i h t-1 ) (4)
f t ＝σ(W f x t +U f h t-1 ) (5)
o t ＝σ(W o x t +U o h t-1 ) (6)
Figure FDA0003990433170000044
Figure FDA0003990433170000045
Figure FDA0003990433170000046
wherein the function
Figure FDA0003990433170000047
To represent
Figure FDA0003990433170000048
i t 、f t 、o t An input gate, a forgetting gate and an output gate at the t-th moment, x t Indicates the input at the t-th time, h t-1 Representing the hidden gate output at time t-1,
Figure FDA0003990433170000049
candidate memory cell representing the t-th time, c t The memory cell at the t-th time is shown,
Figure FDA00039904331700000410
representing dot multiplication, where W and U are trainable matrices, W i 、W f 、W o 、W c Respectively show the input x in the input gate, the forgetting gate, the output gate and the memory unit t Corresponding parameter matrix, U i 、U f 、U o 、U c Respectively shown in the input gate, the forgetting gate, the output gate and the memory unit, and the output h of the hidden gate t-1 The corresponding parameter matrix.
CN202110334529.7A 2021-03-29 2021-03-29 Vein identification method based on time domain short-time and long-time feature fusion Active CN113111744B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110334529.7A CN113111744B (en) 2021-03-29 2021-03-29 Vein identification method based on time domain short-time and long-time feature fusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110334529.7A CN113111744B (en) 2021-03-29 2021-03-29 Vein identification method based on time domain short-time and long-time feature fusion

Publications (2)

Publication Number Publication Date
CN113111744A CN113111744A (en) 2021-07-13
CN113111744B true CN113111744B (en) 2023-02-14

Family

ID=76712501

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110334529.7A Active CN113111744B (en) 2021-03-29 2021-03-29 Vein identification method based on time domain short-time and long-time feature fusion

Country Status (1)

Country Link
CN (1) CN113111744B (en)

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6301375B1 (en) * 1997-04-14 2001-10-09 Bk Systems Apparatus and method for identifying individuals through their subcutaneous vein patterns and integrated system using said apparatus and method
CN108460764A (en) * 2018-03-31 2018-08-28 华南理工大学 The ultrasonoscopy intelligent scissor method enhanced based on automatic context and data
CN109034034A (en) * 2018-07-12 2018-12-18 广州麦仑信息科技有限公司 A kind of vein identification method based on nitrification enhancement optimization convolutional neural networks
WO2019028648A1 (en) * 2017-08-08 2019-02-14 Alibaba Group Holding Limited Processing performance data for machine learning
CN109692005A (en) * 2018-12-03 2019-04-30 南京邮电大学 Personal identification method based on PPG signal sparse decomposition
CN112036316A (en) * 2020-08-31 2020-12-04 中国科学院半导体研究所 Finger vein identification method and device, electronic equipment and readable storage medium
CN112487945A (en) * 2020-11-26 2021-03-12 上海贝业斯健康科技有限公司 Pulse condition identification method based on double-path convolution neural network fusion

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111651941B (en) * 2020-04-30 2022-05-17 北京航空航天大学 Global ionosphere electron total content prediction algorithm
CN112418014B (en) * 2020-11-09 2023-11-17 南京信息工程大学滨江学院 Modulated signal identification method based on wavelet transformation and convolution long-term and short-term memory neural network

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6301375B1 (en) * 1997-04-14 2001-10-09 Bk Systems Apparatus and method for identifying individuals through their subcutaneous vein patterns and integrated system using said apparatus and method
WO2019028648A1 (en) * 2017-08-08 2019-02-14 Alibaba Group Holding Limited Processing performance data for machine learning
CN108460764A (en) * 2018-03-31 2018-08-28 华南理工大学 The ultrasonoscopy intelligent scissor method enhanced based on automatic context and data
CN109034034A (en) * 2018-07-12 2018-12-18 广州麦仑信息科技有限公司 A kind of vein identification method based on nitrification enhancement optimization convolutional neural networks
CN109692005A (en) * 2018-12-03 2019-04-30 南京邮电大学 Personal identification method based on PPG signal sparse decomposition
CN112036316A (en) * 2020-08-31 2020-12-04 中国科学院半导体研究所 Finger vein identification method and device, electronic equipment and readable storage medium
CN112487945A (en) * 2020-11-26 2021-03-12 上海贝业斯健康科技有限公司 Pulse condition identification method based on double-path convolution neural network fusion

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
指脉脉搏波自动识别及自动采集控制方法;杨琳等;《北京工业大学学报》;20130810(第08期);第1261-1263页 *

Also Published As

Publication number Publication date
CN113111744A (en) 2021-07-13

Similar Documents

Publication Publication Date Title
CN106473750B (en) Personal identification method based on photoplethysmographic optimal period waveform
Page et al. Utilizing deep neural nets for an embedded ECG-based biometric authentication system
CN108009520A (en) A kind of finger vein identification method and system based on convolution variation self-encoding encoder neutral net
CN107736894A (en) A kind of electrocardiosignal Emotion identification method based on deep learning
CN110619322A (en) Multi-lead electrocardio abnormal signal identification method and system based on multi-flow convolution cyclic neural network
CN110458197A (en) Personal identification method and its system based on photoplethysmographic
CN110515456A (en) EEG signals emotion method of discrimination and device based on attention mechanism
CN113887513A (en) Motor imagery electroencephalogram signal classification method based on parallel CNN-transform neural network
CN112257518A (en) ECG identity recognition method based on WT and WOA-PNN algorithm
CN109124620B (en) Atrial fibrillation detection method, device and equipment
CN102567708B (en) Method and device for extracting biological feature, biological identification method and system
Patil et al. A non-contact PPG biometric system based on deep neural network
CN111222498A (en) Identity recognition method based on photoplethysmography
Pinto et al. Deep neural networks for biometric identification based on non-intrusive ECG acquisitions
CN109171753A (en) A kind of EEG signals EEG personal identification method encoding neural network certainly based on depth
CN111291727A (en) Method and device for detecting signal quality by photoplethysmography
CN110192864B (en) Cross-domain electrocardiogram biological characteristic identity recognition method
CN114638272A (en) Identity recognition method and device based on fingertip pulse wave signals
CN113243902B (en) Feature extraction method based on photoplethysmography
CN110338791A (en) Surface electromyogram signal active segment dividing method based on Sample Entropy and Gauss model
CN112215196A (en) Electrocardio identity recognition method
CN109919050B (en) Identity recognition method and device
CN113111744B (en) Vein identification method based on time domain short-time and long-time feature fusion
CN113729738A (en) Construction method of multi-channel electromyographic feature image
CN112307996A (en) Fingertip electrocardiogram identity recognition device and method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant